Estimation results

5.1 Spillovers through production activities

We start with estimating equation (1) in which we assume that MNE knowledge spills over through production activities, using the MNE industry labor force aggregate as a proxy for the size of their production activities. Columns (1) and (2) in Table 9 [ PDF 57.1KB | 1 page ] show results from the OLS and the system GMM estimations. As we noted in the previous section, the system GMM approach can correct for possible biases due to the endogeneity of the regressors and firm-specific fixed effects. According to the p value of the Hansen J statistics reported in the last row, we cannot reject the hypothesis that the instruments are orthogonal to the error term at the 5% significance level. This is actually the case in all other estimations presented later. In other words, the use of the lagged regressors as instruments can be justified in all estimations, suggesting that the system GMM is the preferred estimator.

Columns (1) and (2) of Table 9 indicate that although the effect of the industry aggregate of MNE labor is positive and statistically significant in the OLS results, its effect is insignificant in the GMM results. The difference between the OLS and the GMM results suggests that MNEs are more attracted to more productive industries, but that MNE production activities do not improve domestic firm productivity due to knowledge spillovers.

5.2 Differences in FDI spillovers across home countries

In Section 5.1, we assumed that the spillover effect does not vary in size depending on the MNE home country. However, since MNE characteristics substantially differ across home countries, as we have seen in Section 4, this assumption may not hold in practice. Therefore, we estimate the effect of spillovers from Japanese, US, and other MNEs separately, by incorporating the total labor force for each of the three types of MNE in equation (1). The results from the OLS and the GMM estimation are presented in columns (3)–(8) of Table 9. The GMM results indicate that the coefficient on US MNEs is positive and significant, while Japanese and other MNEs do not have a significant effect on local productivity. These results suggest that although Japanese MNE employment does not lead to knowledge spillovers to domestic firms, knowledge spillovers from US MNEs in fact take place.

The size of the spillover effect of non-Japanese MNEs is large. The average across industries of the log of industry labor force of US MNEs increased from 2.97 in 2000 to 3.47 in 2003. Thus, the estimation result in column (6) of Table 9 indicates that the increase in US MNEs raised the average productivity and output of domestic firms by roughly 3.3% (= 0.065 x (3.47 - 2.97)) during the four-year period.

5.3 Spillovers through employment of educated workers

Now, what are reasons for the substantial difference in spillover effects between Japanese and non-Japanese MNEs? In Tables 7 and 8, we found that Japanese MNEs utilize educated labor substantially less than US MNEs. This may be a reason for no spillovers from Japanese MNEs. To test this hypothesis, we employ the MNE industry aggregate of educated workers as an additional proxy for the size of MNEs and estimate equation (1).

Table 10 [ PDF 57.5KB | 1 page ] shows GMM results based on three alternative definitions of educated workers; workers with a master's or higher degree, those with a bachelor's or higher degree, and those with overseas education. In column (1), we find that MNEs' workers with graduate education have a positive and significant impact on the productivity of domestic firms in the same industry, while the effect of total MNE labor force is insignificant as before. However, column (2) indicates that MNE workers with only an undergraduate education do not contribute to productivity improvement of domestic firms. The comparison between the two findings suggest that MNE knowledge spills over through highly educated workers, i.e., holders of master's and doctorates, but not through holders of bachelor's degrees.

In column (3) of Table 10, we find that MNE employment of workers with overseas education also promotes spillovers from FDI. Since the definition of overseas education in our data set includes both graduate and undergraduate education, this result suggests that bachelor's degree holders of foreign universities contribute to spillovers, although bachelor's degree holders from PRC universities do not. This difference is probably due to the multicultural background and linguistic strength of returnees from overseas, which promotes interaction between foreign and domestic workers and thus knowledge spillovers.

5.4 Robustness checks

To check the robustness of the results, we further experiment with three alternative specifications. First, when we constructed firms own R&D stock, we imposed several assumptions as indicated in the Appendix. Since these assumptions may be too strong, we exclude firms' own R&D stock from the set of regressors in the production function equation. Second, following studies such as Javorcik (2004), we first estimate productivity at the firm level using the method developed by Olley and Pakes (1996) and then estimate the impact of FDI variables on the productivity level. The estimated elasticity of capital and labor are 0.232 and 0.743, respectively, which is not very different from the elasticity from the previous system GMM estimations. Third, we use a trans-log production function, rather than a Cobb- Douglas function indicated by equation (1), incorporating the squared log of capital and labor and the interaction between the log of capital and labor in the estimation equation. The results from the three alternative specifications, not shown here for simple presentation but available upon request, are very similar to the results from the benchmark specification presented in Table 10, suggesting the robustness of the benchmark results. In the third alternative specification, the coefficient on the squared terms and the interaction term is always insignificant, and in some cases the coefficient on the log of capital and labor also becomes insignificant, probably due to multicollinearity. Therefore, we do not use the translog function for our benchmark regression.

In addition, since the elasticity of capital and labor may differ across industries, we divided firms into four broadly-defined industry categories,¹² and estimated the elasticity of capital and labor using the method of Olley and Pakes (1996) for each category. For one category, the estimated elasticity of capital exceeded one, whereas for another, the estimated elasticity of labor exceeded one. These values are not acceptable for elasticity of capital or labor. A possible reason for the seemingly biased estimates is that the influence of outliers is exaggerated in smaller subsamples of firms. However, as an experiment, we ignored the possible estimation biases and used the industry-specific elasticity to construct the TFP level for each firm to estimate the spillover effects using the two-step approach of Javorcik (2004). Then, we found that the OLS results are similar to the OLS results in the benchmark estimation assuming the same elasticity of capital and labor across industries. However, in all GMM estimations, the orthogonality between the instruments and the error term was rejected according to the Hansen J test, and hence the instruments were invalid. Therefore, we do not assume discrepancies in the capital and labor elasticity across industries but stick with the benchmark assumption of the same elasticity. One justification of this assumption is that entries to the Z-Park are restricted to high-tech firms so that discrepancies across firms and industries may not be as large as in the case when the sample covers any type of firm in the economy. In addition, we experimented with a trans-log production function, in which we do not assume constant elasticity, and found that the coefficient on the squared terms and the interaction term is insignificant, as we mentioned above. This result also justifies the benchmark assumption.

5.5 Differences in the spillover effect across industries

So far, we have assumed that the spillover effect is the same in size across industries, but the effect may vary depending on industry characteristics. To check this possibility, we incorporated the interaction term between the industry aggregate of educated MNE workers and industry characteristics. In particular, we hypothesize that the size of FDI spillovers through educated labor depends on the importance of R&D activities represented by the ratio of R&D expenditures to sales, or the R&D intensity.

The results are shown in columns (1)–(3) of Table 11 [ PDF 59.2KB | 1 page ], indicating that for all three different definitions of educated labor, the effect of the interaction term is positive and statistically significant, while the industry aggregate of educated MNE workers alone has no significant effect. This evidence suggests that the size of FDI spillovers through educated MNE workers is larger in more R&D-intensive industries.

In addition, we estimate the effect of the interaction term between the industry aggregate of MNE labor force and the industry R&D intensity, in order to check whether we can find knowledge spillovers through MNE production activities under the new assumption. However, as column (4) of Table 11 shows, we find no significant impact for either the interaction term or the industry aggregate labor alone. This evidence suggests that, even if we assume variations in the size of spillover effects across industries, our main conclusion remains unchanged: MNE knowledge spills over through their employment of educated labor, but not through their production alone.

Download this Paper [ PDF 194.5KB| 27 pages ].

Comment(s)

There are [0] comment(s) for this entry. Post a comment.